def __load_model(self):
# 2: Build the Keras model (and possibly load some trained weights)
K.clear_session() # Clear previous models from memory.
# The output `predictor_sizes` is needed below to set up `SSDBoxEncoder`
model, predictor_sizes = ssd_300(
image_size=(self.img_train_height, self.img_train_width,
self.img_channels),
n_classes=self.n_classes,
min_scale=
None, # You could pass a min scale and max scale instead of the `scales` list, but we're not doing that here
max_scale=None,
scales=self.scales,
aspect_ratios_global=None,
aspect_ratios_per_layer=self.aspect_ratios,
two_boxes_for_ar1=self.two_boxes_for_ar1,
limit_boxes=self.limit_boxes,
variances=self.variances,
coords=self.coords,
normalize_coords=self.normalize_coords)
# Set the path to the VGG-16 weights below.
### Set up training
# 3: Instantiate an Adam optimizer and the SSD loss function and compile the model
adam = Adam(lr=0.001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=5e-04)
ssd_loss = SSDLoss(neg_pos_ratio=3, n_neg_min=0, alpha=0.1)
model.compile(optimizer=adam, loss=ssd_loss.compute_loss)
# 4: Instantiate an encoder that can encode ground truth labels into the format needed by the SSD loss function
ssd_box_encoder = SSDBoxEncoder(
img_height=self.img_train_height,
img_width=self.img_train_width,
n_classes=self.n_classes,
predictor_sizes=self.predictor_sizes,
min_scale=None,
max_scale=None,
scales=self.scales,
aspect_ratios_global=None,
aspect_ratios_per_layer=self.aspect_ratios,
two_boxes_for_ar1=self.two_boxes_for_ar1,
limit_boxes=self.limit_boxes,
variances=self.variances,
pos_iou_threshold=0.5,
neg_iou_threshold=0.2,
coords=self.coords,
normalize_coords=self.normalize_coords)
adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=5e-05)
ssd_loss = SSDLoss(neg_pos_ratio=3, n_neg_min=0, alpha=0.1)
model.compile(optimizer=adam, loss=ssd_loss.compute_loss)
# 4: Instantiate an encoder that can encode ground truth labels into the format needed by the SSD loss function
ssd_box_encoder = SSDBoxEncoder(img_height=img_height,
img_width=img_width,
n_classes=n_classes,
predictor_sizes=predictor_sizes,
min_scale=None,
max_scale=None,
scales=scales,
aspect_ratios_global=None,
aspect_ratios_per_layer=aspect_ratios,
two_boxes_for_ar1=two_boxes_for_ar1,
limit_boxes=limit_boxes,
variances=variances,
pos_iou_threshold=0.5,
neg_iou_threshold=0.2,
coords=coords,
normalize_coords=normalize_coords)
# 5: Create the training set batch generator
classes = [
'background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus',
'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike',
'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'
]
num_classes=n_classes)
logging.debug('... Complete!')
# ========== MODEL TRAINING ========== #
if args.train:
# encoder to make ground truth into SSD loss format
logging.debug('Starting Box Encoder')
predictor_sizes = model.predictor_sizes
ssd_box_encoder = SSDBoxEncoder(img_height=img_height,
img_width=img_width,
n_classes=n_classes,
predictor_sizes=predictor_sizes,
scales=[0.08, 0.16, 0.32, 0.64, 0.96],
aspect_ratios_global=[0.5, 1.0, 2.0],
aspect_ratios_per_layer=None,
two_boxes_for_ar1=True,
steps=None,
offsets=None,
limit_boxes=False,
variances=[1.0, 1.0, 1.0, 1.0],
pos_iou_threshold=0.5,
neg_iou_threshold=0.2,
coords='centroids',
normalize_coords=False)
# generator
logging.debug('Making BatchGenerator and defining generator function')
train_dataset = BatchGenerator(
labels_path=args.labels,
box_output_format=['class_id', 'xmin', 'xmax', 'ymin', 'ymax'])
train_generator = train_dataset.generate(batch_size=args.batch_size,
shuffle=True,
def train_hs512(lr=1e-4,
freeze_bn=False,
optim=None,
batch_size=8,
weights_path=None,
save_weights_only=True,
epochs=25):
if weights_path is None:
weights_path = 'VGG_VOC0712Plus_SSD_512x512_ft_iter_160000.h5'
img_height = 512 # 1088 // 2 # Height of the input images
img_width = 512 # 2048 // 2 # Width of the input images
img_channels = 3 # Number of color channels of the input images
# subtract_mean = [104, 117, 123] # The per-channel mean of the images in the dataset
subtract_mean = [138, 138,
138] # The per-channel mean of the images in the dataset
swap_channels = False
n_classes = 20 # The number of positive classes, e.g. 20 for Pascal VOC, 80 for MS COCO.
scales = [0.04, 0.1, 0.26, 0.42, 0.58, 0.74, 0.9, 1.06] # MS COCO scales
# scales = [0.07, 0.15, 0.3, 0.45, 0.6, 0.75, 0.9, 1.05]
aspect_ratios = [
[1.0,
2.0, 0.5], [1.0, 2.0, 0.5, 3.0,
1.0 / 3.0],
[1.0, 2.0, 0.5, 3.0,
1.0 / 3.0],
[1.0, 2.0, 0.5,
3.0, 1.0 / 3.0], [1.0, 2.0, 0.5, 3.0, 1.0 / 3.0
], [1.0, 2.0, 0.5
], [1.0, 2.0, 0.5]
] # The anchor box aspect ratios used in the original SSD300; the order matters
two_boxes_for_ar1 = True
# The space between two adjacent anchor box center points for each predictor layer.
steps = [8, 16, 32, 64, 128, 256, 512]
offsets = [0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5]
limit_boxes = False # Whether or not you want to limit the anchor boxes to lie entirely within the image boundaries
variances = [0.1, 0.1, 0.2, 0.2]
coords = 'centroids'
normalize_coords = True
# 1: Build the Keras model
K.clear_session()
model, pred_sizes = ssd_512(image_size=(img_height, img_width,
img_channels),
n_classes=n_classes,
l2_regularization=0.0005,
scales=scales,
aspect_ratios_per_layer=aspect_ratios,
two_boxes_for_ar1=two_boxes_for_ar1,
steps=steps,
offsets=offsets,
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
subtract_mean=None,
divide_by_stddev=None,
swap_channels=swap_channels,
return_predictor_sizes=True)
# 2: Load the trained VGG-16 weights into the model.
model.load_weights(weights_path, by_name=True)
# 3: Instantiate the optimizer and the SSD loss function and compile the model
if optim is None:
optim = SGD(lr=lr, momentum=0.9)
ssd_loss = SSDLoss(neg_pos_ratio=3, n_neg_min=0, alpha=1.0)
# 4: freeze the base model if needed
if freeze_bn:
for l in model.layers[:38]:
l.trainable = False
# 5: compile model
model.compile(optimizer=optim, loss=ssd_loss.compute_loss)
## Prepare data generation
# 1: Instantiate to `BatchGenerator` objects: One for training, one for validation.
train_dataset = BatchGenerator(
box_output_format=['class_id', 'xmin', 'ymin', 'xmax', 'ymax'])
val_dataset = BatchGenerator(
box_output_format=['class_id', 'xmin', 'ymin', 'xmax', 'ymax'])
# 2: Parse the image and label lists for the training and validation datasets. This can take a while.
images_path_root = './Datasets/'
train_combined_labels = './Datasets/train_combined_ssd_512.txt'
val_labels = './Datasets/val_ssd_512.txt'
train_dataset.parse_csv(images_dir=images_path_root,
labels_filename=train_combined_labels,
input_format=[
'image_name', 'class_id', 'xmin', 'xmax',
'ymin', 'ymax'
])
val_dataset.parse_csv(images_dir=images_path_root,
labels_filename=val_labels,
input_format=[
'image_name', 'class_id', 'xmin', 'xmax', 'ymin',
'ymax'
])
# 3: Instantiate an encoder that can encode ground truth labels into the format needed by the SSD loss function.
ssd_box_encoder = SSDBoxEncoder(img_height=img_height,
img_width=img_width,
n_classes=n_classes,
predictor_sizes=pred_sizes,
min_scale=None,
max_scale=None,
scales=scales,
aspect_ratios_global=None,
aspect_ratios_per_layer=aspect_ratios,
two_boxes_for_ar1=two_boxes_for_ar1,
steps=steps,
offsets=offsets,
limit_boxes=limit_boxes,
variances=variances,
pos_iou_threshold=0.5,
neg_iou_threshold=0.2,
coords=coords,
normalize_coords=normalize_coords)
# 4: Set the image processing / data augmentation options and create generator handles.
train_generator = train_dataset.generate(
batch_size=batch_size,
shuffle=True,
train=True,
ssd_box_encoder=ssd_box_encoder,
equalize=False,
brightness=(0.5, 2, 0.5),
flip=0.5,
translate=False,
scale=False,
max_crop_and_resize=(img_height, img_width, 1, 3),
# This one is important because the Pascal VOC images vary in size
random_pad_and_resize=(img_height, img_width, 1, 3, 0.5),
# This one is important because the Pascal VOC images vary in size
random_crop=False,
crop=False,
resize=False,
gray=True,
limit_boxes=True,
# While the anchor boxes are not being clipped, the ground truth boxes should be
include_thresh=0.4)
val_generator = val_dataset.generate(
batch_size=batch_size,
shuffle=True,
train=True,
ssd_box_encoder=ssd_box_encoder,
equalize=False,
brightness=False,
flip=False,
translate=False,
scale=False,
max_crop_and_resize=(img_height, img_width, 1, 3),
# This one is important because the Pascal VOC images vary in size
random_pad_and_resize=(img_height, img_width, 1, 3, 0.5),
# This one is important because the Pascal VOC images vary in size
random_crop=False,
crop=False,
resize=False,
gray=True,
limit_boxes=True,
include_thresh=0.4)
# Get the number of samples in the training and validations datasets to compute the epoch lengths below.
n_train_samples = train_dataset.get_n_samples()
n_val_samples = val_dataset.get_n_samples()
# ## 4. Run the training
fingerprint = 'ssd{:%Y-%m-%d_%H-%M-%S}'.format(datetime.datetime.now())
tbCallBack = TensorBoard(log_dir='./Graph/{}'.format(fingerprint),
histogram_freq=0,
batch_size=batch_size,
write_graph=True)
checkpointer = ModelCheckpoint(
'./saved/{{val_loss:.2f}}__{}_best_weights.h5'.format(fingerprint),
monitor='val_loss',
verbose=1,
save_best_only=save_weights_only,
save_weights_only=True,
mode='auto',
period=1)
learning_rate_reduction = ReduceLROnPlateau(monitor='val_loss',
patience=5,
verbose=1,
factor=0.5,
min_lr=1e-6)
stopper = EarlyStopping(monitor='val_loss', min_delta=0.001, patience=10)
epochs = 30
history = model.fit_generator(
generator=train_generator,
steps_per_epoch=ceil(n_train_samples / batch_size),
epochs=epochs,
callbacks=[checkpointer, learning_rate_reduction, stopper, tbCallBack],
validation_data=val_generator,
validation_steps=ceil(n_val_samples / batch_size))
def train(args):
model = ssd_300(mode='training',
image_size=(img_height, img_width, img_channels),
n_classes=n_classes,
l2_regularization=0.0005,
scales=scales,
aspect_ratios_per_layer=aspect_ratios,
two_boxes_for_ar1=two_boxes_for_ar1,
steps=steps,
offsets=offsets,
limit_boxes=limit_boxes,
variances=variances,
coords=coords,
normalize_coords=normalize_coords,
subtract_mean=subtract_mean,
divide_by_stddev=None,
swap_channels=swap_channels)
model.load_weights(args.weight_file, by_name=True, skip_mismatch=True)
predictor_sizes = [
model.get_layer('conv11_mbox_conf').output_shape[1:3],
model.get_layer('conv13_mbox_conf').output_shape[1:3],
model.get_layer('conv14_2_mbox_conf').output_shape[1:3],
model.get_layer('conv15_2_mbox_conf').output_shape[1:3],
model.get_layer('conv16_2_mbox_conf').output_shape[1:3],
model.get_layer('conv17_2_mbox_conf').output_shape[1:3]
]
adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=5e-04)
ssd_loss = SSDLoss(neg_pos_ratio=3, n_neg_min=0, alpha=1.0)
model.compile(optimizer=adam,
loss=ssd_loss.compute_loss,
metrics=['accuracy'])
train_dataset = BatchGenerator(
box_output_format=['class_id', 'xmin', 'ymin', 'xmax', 'ymax'])
val_dataset = BatchGenerator(
box_output_format=['class_id', 'xmin', 'ymin', 'xmax', 'ymax'])
# 2: Parse the image and label lists for the training and validation datasets. This can take a while.
# TODO: Set the paths to the datasets here.
COCO_format_val_images_dir = args.ms_coco_dir_path + '/val/'
COCO_format_train_images_dir = args.ms_coco_dir_path + '/train/'
COCO_format_train_annotation_dir = args.ms_coco_dir_path + '/annotations/train.json'
COCO_format_val_annotation_dir = args.ms_coco_dir_path + '/annotations/val.json'
VOC_2007_images_dir = args.voc_dir_path + '/VOC2007/JPEGImages/'
VOC_2012_images_dir = args.voc_dir_path + '/VOC2012/JPEGImages/'
# The directories that contain the annotations.
VOC_2007_annotations_dir = args.voc_dir_path + '/VOC2007/Annotations/'
VOC_2012_annotations_dir = args.voc_dir_path + '/VOC2012/Annotations/'
# The paths to the image sets.
VOC_2007_train_image_set_filename = args.voc_dir_path + '/VOC2007/ImageSets/Main/trainval.txt'
VOC_2012_train_image_set_filename = args.voc_dir_path + '/VOC2012/ImageSets/Main/trainval.txt'
VOC_2007_val_image_set_filename = args.voc_dir_path + '/VOC2007/ImageSets/Main/test.txt'
# The XML parser needs to now what object class names to look for and in which order to map them to integers.
classes = [
'background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus',
'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse',
'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train',
'tvmonitor'
]
'''
This is an JSON parser for the MS COCO datasets. It might be applicable to other datasets with minor changes to
the code, but in its current form it expects the JSON format of the MS COCO datasets.
Arguments:
images_dirs (list, optional): A list of strings, where each string is the path of a directory that
contains images that are to be part of the dataset. This allows you to aggregate multiple datasets
into one (e.g. one directory that contains the images for MS COCO Train 2014, another one for MS COCO
Val 2014, another one for MS COCO Train 2017 etc.).
annotations_filenames (list): A list of strings, where each string is the path of the JSON file
that contains the annotations for the images in the respective image directories given, i.e. one
JSON file per image directory that contains the annotations for all images in that directory.
The content of the JSON files must be in MS COCO object detection format. Note that these annotations
files do not necessarily need to contain ground truth information. MS COCO also provides annotations
files without ground truth information for the test datasets, called `image_info_[...].json`.
ground_truth_available (bool, optional): Set `True` if the annotations files contain ground truth information.
include_classes (list, optional): Either 'all' or a list of integers containing the class IDs that
are to be included in the dataset. Defaults to 'all', in which case all boxes will be included
in the dataset.
ret (bool, optional): Whether or not the image filenames and labels are to be returned.
Returns:
None by default, optionally the image filenames and labels.
'''
train_dataset.parse_json(
images_dirs=[COCO_format_train_images_dir],
annotations_filenames=[COCO_format_train_annotation_dir],
ground_truth_available=True,
include_classes='all',
ret=False)
val_dataset.parse_json(
images_dirs=[COCO_format_val_images_dir],
annotations_filenames=[COCO_format_val_annotation_dir],
ground_truth_available=True,
include_classes='all',
ret=False)
# 3: Instantiate an encoder that can encode ground truth labels into the format needed by the SSD loss function.
ssd_box_encoder = SSDBoxEncoder(img_height=img_height,
img_width=img_width,
n_classes=n_classes,
predictor_sizes=predictor_sizes,
min_scale=None,
max_scale=None,
scales=scales,
aspect_ratios_global=None,
aspect_ratios_per_layer=aspect_ratios,
two_boxes_for_ar1=two_boxes_for_ar1,
steps=steps,
offsets=offsets,
limit_boxes=limit_boxes,
variances=variances,
pos_iou_threshold=0.5,
neg_iou_threshold=0.2,
coords=coords,
normalize_coords=normalize_coords)
batch_size = args.batch_size
train_generator = train_dataset.generate(
batch_size=batch_size,
shuffle=True,
train=True,
ssd_box_encoder=ssd_box_encoder,
convert_to_3_channels=True,
equalize=False,
brightness=(0.5, 2, 0.5),
flip=0.5,
translate=False,
scale=False,
max_crop_and_resize=(img_height, img_width, 1, 3),
# This one is important because the Pascal VOC images vary in size
random_pad_and_resize=(img_height, img_width, 1, 3, 0.5),
# This one is important because the Pascal VOC images vary in size
random_crop=False,
crop=False,
resize=False,
gray=False,
limit_boxes=True,
# While the anchor boxes are not being clipped, the ground truth boxes should be
include_thresh=0.4)
val_generator = val_dataset.generate(
batch_size=batch_size,
shuffle=True,
train=True,
ssd_box_encoder=ssd_box_encoder,
convert_to_3_channels=True,
equalize=False,
brightness=(0.5, 2, 0.5),
flip=0.5,
translate=False,
scale=False,
max_crop_and_resize=(img_height, img_width, 1, 3),
# This one is important because the Pascal VOC images vary in size
random_pad_and_resize=(img_height, img_width, 1, 3, 0.5),
# This one is important because the Pascal VOC images vary in size
random_crop=False,
crop=False,
resize=False,
gray=False,
limit_boxes=True,
# While the anchor boxes are not being clipped, the ground truth boxes should be
include_thresh=0.4)
# Get the number of samples in the training and validations datasets to compute the epoch lengths below.
n_train_samples = train_dataset.get_n_samples()
n_val_samples = val_dataset.get_n_samples()
def lr_schedule(epoch):
if epoch <= 300:
return 0.001
else:
return 0.0001
learning_rate_scheduler = LearningRateScheduler(schedule=lr_schedule)
checkpoint_path = args.checkpoint_path + "/ssd300_epoch-{epoch:02d}.h5"
checkpoint = ModelCheckpoint(checkpoint_path)
log_path = args.checkpoint_path + "/logs"
tensorborad = TensorBoard(log_dir=log_path,
histogram_freq=1,
write_graph=True,
write_images=False)
callbacks = [checkpoint, tensorborad, learning_rate_scheduler]
# TODO: Set the number of epochs to train for.
epochs = args.epochs
intial_epoch = args.intial_epoch
history = model.fit_generator(
generator=train_generator,
steps_per_epoch=ceil(n_train_samples) / batch_size,
verbose=1,
initial_epoch=intial_epoch,
epochs=epochs,
validation_data=val_generator,
validation_steps=ceil(n_val_samples) / batch_size,
callbacks=callbacks)